Crowd Counting Algorithm Based On Deep Convolutional Neural Network

With the advance of the global urbanization process,intelligent monitoring has become a research hotspot in the field of computer vision.As one of the core problems of intelligent monitoring,crowd counting is of great significance in applications such as crowd limiting and crowd drainage.At present,great progress has been made in the research work of crowd counting.However,it is still a big challenge to solve the problem of inconsistency in the image scale of pedestrians under different scenarios.In recent years,deep convolutional neural networks?CNN?has made great achievements in the field of computer vision.Due to its outstanding performances in image feature extraction and model generalization,deep CNN solves the problem of crowd image features extraction in complex backgrounds much more effectively.Existing CNN-based crowd counting methods often apply multi-column or multi-network model to solve this kind of problems,which are more complicated for optimization and computation wasting.In order to extract the scale-relevant features on single image crowd counting task,we proposed a deep residual convolutional neural networks?Res-CNN?for crowd counting based on the deep residual learning theory.As deep residual neural networks can be equivalent to a variety of shallow networks ensemble learning,it can be used to extract scale-relevant crowd image features.Compared with the multi-columns or multi-networks structure,Res-CNN achieved cross-layers information sharing by the combination of “short-cut”,increasing the number of feature scales.Experiments on Shanghaitech,UCF_CC₅0,WorldExp'10 and UCSD datasets show that Res-CNN achieves a similar accuracy and robustness to state-of-the-art methods with a significant reduction in the number of parameters,which saves the computing resources on monitoring devices.We also proposed a multi-scale convolutional neural networks?MSCNN?based on multi-scale convolution blob?MSB Conv?that composed of various sizes of kernel.By filling feature maps with relative edge padding before multi-scale convolution,the MSB Conv achieved inner-layers information sharing and greatly reduces the redundancy parameters.Like Res-CNN,MSCNN is an end-to-end single-column convolutional neural networks with low training complexity.Experiments on the 4 crowd counting datasets mentioned above proved that MSCNN outperforms the state-of-the-art methods on both accuracy and robustness with far less number of parameters,which is more computing saving.In summary,this paper proposes two kinds of convolutional neural networks.They are applied to extract multi-scales feature by using vertical and horizontal direction for crowd counting instead of using single-column structure.This work will broad the fields in related research in CNN-based intelligent monitoring.